On Mon, 3 Aug 2015 20:50:43 +0800 Baoquan He <bhe at redhat.com> wrote:
> People reported that crash_notes in /proc/vmcore were corrupted and
> this cause crash kdump failure. With code debugging and log we got
> the root cause. This is because percpu variable crash_notes are
> allocated in 2 vmalloc pages. Currently percpu is based on vmalloc
> by default. Vmalloc can't guarantee 2 continuous vmalloc pages
> are also on 2 continuous physical pages. So when 1st kernel exports
> the starting address and size of crash_notes through sysfs like below:
>
> /sys/devices/system/cpu/cpux/crash_notes
> /sys/devices/system/cpu/cpux/crash_notes_size
>
> kdump kernel use them to get the content of crash_notes. However the
> 2nd part may not be in the next neighbouring physical page as we
> expected if crash_notes are allocated accross 2 vmalloc pages. That's
> why nhdr_ptr->n_namesz or nhdr_ptr->n_descsz could be very huge in
> update_note_header_size_elf64() and cause note header merging failure
> or some warnings.
>
> In this patch change to call __alloc_percpu() to passed in the align
> value by rounding crash_notes_size up to the nearest power of two.
> This make sure the crash_notes is allocated inside one physical page
> since sizeof(note_buf_t) in all ARCHS is smaller than PAGE_SIZE.
> Meanwhile add a WARN_ON in case it grows to be bigger than PAGE_SIZE
> in the future.
>
> ...
>
> --- a/kernel/kexec.c
> +++ b/kernel/kexec.c
> @@ -1620,7 +1620,16 @@ void crash_save_cpu(struct pt_regs *regs, int cpu)
> static int __init crash_notes_memory_init(void)
> {
> /* Allocate memory for saving cpu registers. */
> - crash_notes = alloc_percpu(note_buf_t);
> + size_t size, align;
> + int order;
> +
> + size = sizeof(note_buf_t);
> + order = get_count_order(size);
> + align = min_t(size_t, (1<<order), PAGE_SIZE);
> +
> + WARN_ON(size > PAGE_SIZE);
> +
> + crash_notes = __alloc_percpu(size, align);
A code comment would be helpful - the reason for this code's existence
is otherwise utterly unobvious.
I think it can be done this way:
align = min(roundup_pow_of_two(sizeof(note_buf_t)), PAGE_SIZE);
I never noticed get_count_order() before. afaict it does the same as
order_base_2(), except get_count_order() generates better code and has
a ridiculous name.
And I think the WARN_ON can be replaced with a
BUILD_BUG_ON(sizeof>PAGE_SIZE)? That would avoid adding runtime
overhead.
